Transfer Learning with MobileNetV3 for High-Accuracy Classification of Tomato Plant Diseases

Abstract

This research proposes a deep learning based transfer learning approach for automatic identification of tomato plant diseases. The system utilizes MobileNetV3Large as a backbone for feature extraction and is further fine-tuned on a custom dataset based on the Plant Village dataset with total of 14,529 images divided over 10 classes (nine disease classes and one healthy class). The method used for training was a two-step approach where we first froze the base network and trained a new classifier head and then we fine-tuned the last 40 layers of the feature extractor. Data augmentation techniques such as rotation and horizontal flipping were employed during training to facilitate generalization. We present experimental results that show a remarkable classification performance, with a test set accuracy of 98.91% with a loss of 0.0361. When the fine-tuning happened it produced a validation accuracy of 97.35% This study validates MobileNetV3 as a powerful light-weight backbone for real-time, vision-based diagnosis of plant diseases, making a promising contribution to precision agriculture applications.